![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3574256%2Fab8342a27c6c35cc592af1667f2d28c8%2Flogo.png?generation=1670575259766962&alt=media\")
This work extends my previous work but uses NoROI (whole image) and KerasCV (augmentation). KerasCV is a cool tool, it kinda reminds of Albumentation. Unlike my previous work these notebooks uses square image for training. From initial experiments, it seems rectangle + roi performs better than square + noroi images. But it comes with a cost of processing time. Need to fine-tune for this work for better comparison.
\nSimilarly to previous work, notebook supports remote-tpu, local-tpu, multi-gpu, and single-gpu training. It also tracks training and stores grad-cam images in Weights & Biases
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3574256%2Feb0c4273221dce28126fa239eb65cac2%2Fbefore-kerascv.png?generation=1670575642614816&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3574256%2Fc9dfcf6e4db706d91d89d102dcc8c79f%2Fafter-kerascv.png?generation=1670575666182059&alt=media\")
You can track all my experiments here
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3574256%2F918e2c409dc4e7ea7dbd346afe886df2%2Fwandb-kerascv.PNG?generation=1670575935756383&alt=media\")
![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3574256%2F69df6baa51eb2a9d5dbdf9ecb15ee653%2Fkerascv-gradcam.png?generation=1670575962505799&alt=media\")
\n", "messages": [ { "id": 2059806, "postDate": "2022-12-09T08:44:33.503Z", "content": "Note: KerasCV requires
\ntf>=2.9hence can't use it during inference. Some augmentations such aszoom,translation,rotationthrows error, perhaps due to implicit dimension intf.data.Dataset
This work extends my previous work but uses NoROI (whole image) and KerasCV (augmentation). KerasCV is a cool tool, it kinda reminds of Albumentation. Unlike my previous work these notebooks uses square image for training. From initial experiments, it seems rectangle + roi performs better than square + noroi images. But it comes with a cost of processing time. Need to fine-tune for this work for better comparison.
\nSimilarly to previous work, notebook supports remote-tpu, local-tpu, multi-gpu, and single-gpu training. It also tracks training and stores grad-cam images in Weights & Biases
You can track all my experiments here
\n
\n", "rawMarkdown": "Note: KerasCV requires
\ntf>=2.9hence can't use it during inference. Some augmentations such aszoom,translation,rotationthrows error, perhaps due to implicit dimension intf.data.Dataset
\n\nThis work extends my [previous](https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/370689) work but uses NoROI (whole image) and [**KerasCV**](https://github.com/keras-team/keras-cv/) (augmentation). KerasCV is a cool tool, it kinda reminds of Albumentation. Unlike my previous work these notebooks uses **square** image for training. From initial experiments, it seems **rectangle + roi** performs better than **square + noroi** images. But it comes with a cost of **processing time**. Need to fine-tune for this work for better comparison.\n\nSimilarly to previous work, notebook supports `remote-tpu`, `local-tpu`, `multi-gpu`, and `single-gpu` training. It also tracks training and stores grad-cam images in [**Weights & Biases**](https://wandb.ai/awsaf49/rsna-bcd-public)\n\n\n## Notebooks\n* train: [RSNA-BCD: NoROI KerasCV [TF][Train]](https://www.kaggle.com/awsaf49/rsna-bcd-noroi-kerascv-tf-train/)\n* infer: [RSNA-BCD: NoROI KerasCV [TF][Infer]](https://www.kaggle.com/awsaf49/rsna-bcd-noroi-kerascv-tf-infer/)\n\n## Before KerasCV Aug.\n\n\n## After KerasCV Aug.\n\n\n## WandB\nYou can track all my experiments [here](https://wandb.ai/awsaf49/rsna-bcd-public)\n\n\n## Grad-CAM\n\n\n> **Note**: KerasCV requires `tf>=2.9` hence can't use it during inference. Some augmentations such as `zoom`, `translation`, `rotation` throws error, perhaps due to implicit dimension in `tf.data.Dataset`",
"votes": 11
}
],
"comments": [],
"raw_markdown_by_id": {
"2059806": "\n\nThis work extends my [previous](https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/370689) work but uses NoROI (whole image) and [**KerasCV**](https://github.com/keras-team/keras-cv/) (augmentation). KerasCV is a cool tool, it kinda reminds of Albumentation. Unlike my previous work these notebooks uses **square** image for training. From initial experiments, it seems **rectangle + roi** performs better than **square + noroi** images. But it comes with a cost of **processing time**. Need to fine-tune for this work for better comparison.\n\nSimilarly to previous work, notebook supports `remote-tpu`, `local-tpu`, `multi-gpu`, and `single-gpu` training. It also tracks training and stores grad-cam images in [**Weights & Biases**](https://wandb.ai/awsaf49/rsna-bcd-public)\n\n\n## Notebooks\n* train: [RSNA-BCD: NoROI KerasCV [TF][Train]](https://www.kaggle.com/awsaf49/rsna-bcd-noroi-kerascv-tf-train/)\n* infer: [RSNA-BCD: NoROI KerasCV [TF][Infer]](https://www.kaggle.com/awsaf49/rsna-bcd-noroi-kerascv-tf-infer/)\n\n## Before KerasCV Aug.\n\n\n## After KerasCV Aug.\n\n\n## WandB\nYou can track all my experiments [here](https://wandb.ai/awsaf49/rsna-bcd-public)\n\n\n## Grad-CAM\n\n\n> **Note**: KerasCV requires `tf>=2.9` hence can't use it during inference. Some augmentations such as `zoom`, `translation`, `rotation` throws error, perhaps due to implicit dimension in `tf.data.Dataset`"
}
}